Genetic Variants Associated With Glycine Metabolism and Their Role in Insulin Sensitivity and Type 2 Diabetes

We analyzed nondiabetic participants of European ancestry from four studies who provided DNA for genome-wide genotyping and underwent a direct measure of insulin sensitivity. The four studies include RISC (n = 957), Botnia (n = 341), EUGENE2 consortium (European network on Functional Genomics of type 2 diabetes; n = 577) and the Stanford Insulin Suppression Test (IST) Cohort (n = 263). The descriptive characteristics of the RISC participants are shown in Table 1. In brief, we excluded individuals with cryptic relatedness using PLINK pairwise identity by descent estimation (PI_HAT >0.2). We excluded individuals with lipid disorders or diabetes, lipid medications, pregnancy, fasting plasma glucose ≥7.0 mmol/L, or 2-h plasma glucose (on a 75-g oral glucose tolerance test) ≥11.0 mmol/L. The individual study characteristics, genotyping, and phenotyping details are provided in Supplementary Data. We performed genome-wide association studies (GWAS) for metabolites in the RISC study, replicated the GWAS findings in the Botnia study, and carried out Mendelian randomization analyses in RISC, EUGENE2, and Stanford IST to test the associations of genetic variants with insulin sensitivity.

We selected 14 metabolites for GWAS. The metabolites were selected based on the study of Gall et al. (2010) (4). We selected metabolites that were both available in the RISC study and associated with insulin sensitivity (3,5). Details are shown in Table 2.

We selected metabolite ratios from the 14 metabolites based on two criteria: 1) the two metabolites were linked by one-step enzymatic reactions, and 2) the ratio was associated with insulin sensitivity measured by hyperinsulinemic-euglycemic clamp (M value). The glycine-to-serine ratio was the only one that satisfied both criteria (Fig. 2D). In both RISC and Botnia studies, metabolites were measured using multiple-platform mass spectrometry technology (ultra-high performance liquid chromatography and gas chromatography) as previously described (17–19). Absolute quantitation was performed for the 14 metabolites (Table 2) for the RISC study samples by UHPLC-MS/MS analysis (4,49).

The plasma concentrations of metabolites were fitted in a linear regression model with adjustment for age, sex, and centers. Then, the standardized residuals were normalized by inverse-normal transformation prior to GWAS. We performed GWAS with each metabolite using MACH2QTL based on an additive genetic model (20,21).

For the glycine-to-serine ratio, we log 10 transformed the ratio and then adjusted for age, sex, and center in linear regression analyses. We performed GWAS using MACH2QTL as with single metabolites’ concentrations.

Some of the key enzymes and transporters involved in the metabolism and transport of metabolites are known. We selected 34 genes for the fourteen metabolites, consisting of carrier-encoding or enzyme-encoding genes involved in the rate-limiting steps of the relevant biosynthetic pathways. The genes selected are listed in Table 2. We classified single nucleotide polymorphisms (SNPs) within 300 kb of these genes as candidate SNPs. To prioritize SNPs for follow-up, we corrected for multiple testing of the total number of SNPs in each candidate region (a conservative threshold, given the correlation between SNPs). However, we still used P value <3 × 10⁻⁹ (5 × 10⁻⁸ corrected for 15 tests [14 metabolites and one ratio]) in all available studies as the final criteria for association.

We used two statistical thresholds to select SNPs for replication. First, we used P value <5 × 10⁻⁸ as the standard for genome-wide significance in the context of common SNPs. Second, for the 34 candidate genes (Table 2 and research design and methods), we divided 0.05 by the total number of SNPs in the gene ±300 kb. We meta-analyzed SNP-metabolite results from RISC and Botnia using an inverse variance–weighted approach as implemented in STATA command “metan.” For Mendelian randomization analyses, we used SNPs reaching P value <3 × 10⁻⁹ in the meta-analysis of the two studies.

To validate SNPs associated with the glycine-to-serine ratio, we linked results from the recently published Cooperative Health Research in the Region of Augsburg (KORA) and UKtwins studies (22) to our GWAS results. We meta-analyzed our results with those from the KORA or UKtwins studies with a significant threshold of P value <3 × 10⁻⁹ when including all available studies.

We performed further analyses for the five SNPs associated with metabolites in the glycine or glutathione biosynthesis pathways, which include glycine, serine, betaine, α-HB, α-KB, and glycine-to-serine ratio. We tested the associations of each SNP against the other metabolite traits. We performed association analyses in the linear regression model described in method section 3 in STATA (version 10.1).

We tested the role of metabolite-associated SNPs reaching genome-wide significance with two diabetes-related traits: hyperinsulinemic-euglycemic clamp (M value corrected for kilograms body weight), which was a measure of whole-body insulin sensitivity, and fasting insulin. M value–based measures of insulin sensitivity were corrected for age, sex, and center and converted to SD units; and inverse normalized. Fasting insulin was natural log transformed; corrected for age, sex, and center; and converted to SD units. Using RISC data, we calculated two estimates of the association between metabolite SNPs and diabetes-related traits for each metabolite trait. First, we calculated an estimated expected effect if there was a causal association between metabolites and diabetes-related measures, using a triangulation approach as shown in Supplementary Fig. 1: we calculated the correlation between standardized metabolite levels and the two diabetes-related traits. We then multiplied these standardized effects by that between metabolite SNPs and metabolites to estimate an approximate expected effect size of the association between metabolite SNPs and M value and fasting insulin. We calculated approximate expected 95% CIs based on the observed effects and SEs using the Taylor series expansion of the ratio of two means (23). Second, we tested the observed effect between metabolite SNPs and the two diabetes-related traits. For the clamp-based measures of insulin sensitivity, we used three studies (RISC, EUGENE2, and Stanford IST) and meta-analyzed results using the program METAL (24). In EUGENE2, insulin sensitivity was measured using the same hyperinsulinemic-euglycemic clamp–based protocol as that used by RISC (25). In the Stanford study, insulin sensitivity was measured by steady-state plasma glucose method. The steady-state plasma glucose value is highly inversely correlated to M value (r = −0.93, P < 0.001) (26), so meta-analyses were performed between the three studies by reversing the signs of the effect sizes in Stanford.

For fasting insulin measures of insulin sensitivity, we used data from the Meta-Analyses of Glucose and Insulin-related traits Consortium (MAGIC), consisting of a meta-analysis of 23 GWAS with 27,589 individuals (27).

We used data from the Diabetes Genetics Replication and Meta-analysis Consortium (DIAGRAM) to assess the association of metabolite SNPs with type 2 diabetes. These data come from 8,130 patients with type 2 diabetes and 38,987 control subjects from eight GWAS (28).

In this RISC study, where we had measures of SNPs, metabolites, and insulin sensitivity, we performed instrumental variable analyses using the two-stage least squares regression approach implemented in the STATA command “ivreg2.” Using the two-stage least squares regression approach, the instrumental variable estimator β_IV provides an estimate of the causal effects of exposure (i.e., metabolites) on outcome (i.e., insulin sensitivity) even in the presence of unmeasured confounders (10,29).

In the RISC study, we identified eight signals of interest either at genome-wide significance or reaching a locus-wide nominal level of significance around one of the candidate genes (Supplementary Table 2). One of these signals represented a widely reported association between SNPs in the FADs gene cluster and fatty acids (27,30,31) (in our case adrenate), and we did not pursue this association further. We successfully genotyped SNPs representing six of the remaining seven signals in the Botnia study (Table 3).

After meta-analysis of RISC and Botnia data (where available), we identified four association signals with three separate single metabolites—the amino acids glycine, serine, and betaine—at P value <3 × 10⁻⁹. For ratios of metabolites, we identified one signal for glycine-to-serine ratio that when meta-analyzed with published KORA data reached P value <3 ×10⁻⁹. Details of the associations are given in Table 3 and Fig. 1. Two SNPs associated with serine were taken forward from the RISC GWAS but did not replicate in the Botnia study (Table 3).

We identified an association between rs499368 in the SLC6A12 gene and betaine levels (P value 1.46 × 10⁻¹⁰) (Fig. 1A). This signal has not previously been reported with any other trait and was not captured at r² > 0.8 in the published KORA or UKtwins data. The second association occurred between rs17823642 near a candidate gene, BHMT, and betaine levels (P value 2.3 × 10⁻⁹) (Fig. 1B). This signal has not previously been reported with any other trait at genome-wide significance but was captured at r² = 1.0 by rs7732845 in the published KORA data (but not UKtwins), and a meta-analysis of RISC, Botnia, and KORA data (P value for KORA alone 1.98 × 10⁻⁶) confirms very robust evidence of association (meta-analyzed P value 6.07 × 10⁻¹⁴).

We identified an association between rs715 in the 3′ untranslated region of the CPS1 gene and glycine levels at genome-wide significance (P value 3.30 × 10⁻⁵⁰) (Fig. 1C). This signal was not captured at r² >0.8 in the published KORA or UKtwins data, but a SNP (rs4673558) with an r² = 0.21 with rs715 is associated with glycine levels with P value 4.3 × 10⁻¹¹ in the UKtwins data.

We identified an association between rs478093 near the PHGDH gene and serine levels at genome-wide significance (P value 1.52 × 10⁻⁹) (SNP not available in Botnia study) (Fig. 1D). This signal was previously reported as associated with serine and ratios of metabolites involving serine in the published KORA and UKtwins data (based on rs477992 [r² = 0.93], meta-analyzed P value 1.94 × 10⁻¹⁴) (22).

We identified a previously unreported association between rs1107366 near the ALDH1L1 gene and glycine-to-serine ratios (P value 2.25 × 10⁻⁶) (Fig. 1E). This signal reached genome-wide significance in combination with data from the KORA and UKtwins studies (meta-analyzed P value 2.8 × 10⁻¹²) (Table 3).

The SNPs in the CPS1 locus have been previously reported with sex-specific effect on glycine and homocysteine levels (rs7422339, r² = 0.92 with rs715) (32,33). We observed a similar sex-specific association between this signal and glycine levels (Supplementary Fig. 2). The association was weak in males (β = −0.19 [95% CI −0.31 to −0.08]; P value = 1.1 × 10⁻³) but more than four times the effect size in females (β = −0.84 [−0.98 to −0.69]; P value 4.5 × 10⁻²⁸). The Z test for the null hypothesis of no sex-specific effect was rejected at P value 2.67 × 10⁻¹³. This result was consistent with the female-specific association previously reported (32,33). There was no evidence that the association was different between pre- and postmenopausal women (premenopausal, effect −0.78 [−0.95 to −0.61]; P value 1.57 × 10⁻¹⁹; postmenopausal, −1.04 [−1.35 to −0.73]; P value 4.57 × 10⁻¹¹). We did not observe any evidence of sex-specific effects for the other metabolite SNPs.

The effects of the five confirmed signals on the other four single metabolites levels and glycine-to-serine ratio are shown in Table 4, with the last column showing the effects of metabolite–insulin sensitivity associations.

Associations between the five metabolite-associated SNPs, metabolite levels, and diabetes-related traits (fasting insulin and hyperinsulinemic-euglycemic clamp [M value]) are shown in Table 5. There were strong correlations between fasting insulin and metabolite levels (0.10 < r < 0.20), as was expected, given that the metabolites were selected as those correlated with a measure of insulin sensitivity. These strong correlations between phenotypes meant that the MAGIC data had >90% power to detect associations at P value < 0.01 based on the estimated expected effects between metabolite SNPs and fasting insulin. However, there were no associations between metabolite SNPs and fasting insulin. The effect sizes observed in the MAGIC data were all smaller than those expected based on the triangulation calculations.

We identified a nominal association between the glycine-to-serine ratio–associated SNP rs1107366 near ALDH1L1 and clamp-based measures of insulin sensitivity (β = 0.09 SD [95% CI 0.03–0.15], where the allele that raises glycine-to-serine ratios increases clamp-based insulin sensitivity P value 0.005) (Table 5 and Supplementary Table 3). The observed effect size on insulin sensitivity (IS) was larger than the expected effect (expected β_SNP-IS = 0.03 SD [95% CI 0.01–0.04]). Results of instrumental variable analyses in RISC were consistent with the main results: the glycine-to-serine ratio predicted from the rs1107366 genotypes is associated with clamp-based insulin sensitivity (β_IV = 1.00 [0.24–1.76]; P value 0.01). No other metabolite SNPs were associated with clamp-based measures of insulin sensitivity in either the triangulation analyses or the instrumental variable analyses, including the other glycine and serine signals (rs715 and rs478093).

There was no evidence of association between four of the five metabolite SNPs and type 2 diabetes, based on the meta-analysis of case-control studies reported by the DIAGRAM study (Supplementary Table 4). The rs715 SNP in CPS1 associated with glycine levels was poorly captured in the type 2 diabetes GWAS meta-analysis (28).

Using a genome-wide approach, we have identified five associations between genetic variants and circulating levels of three metabolites and one metabolite ratio. These metabolites occur in pathways strongly correlated with the gold standard measure of insulin sensitivity (hyperinsulinemic-euglycemic clamp) in the RISC study—primarily, the glycine biosynthesis pathway. Three of these associations have been previously identified at genome-wide levels of significance. We have tested the association of these variants with insulin sensitivity using the largest collective set of studies with these measures, including RISC, EUGENE2, and Stanford.

Glycine and serine are glycogenic amino acids involved in hepatic gluconeogenesis and glutathione biosynthesis, which are potentially important pathways in diabetes and insulin resistance. In the RISC study, glycine, serine, and betaine were positively correlated with clamp-based measures of insulin sensitivity and negatively correlated with fasting insulin. These associations are in line with previous findings (7–9,34).

Using Mendelian randomization analyses, we assessed whether these amino acids play a causal role in insulin sensitivity and type 2 diabetes risks. Our Mendelian randomization analyses do not support a causal association between genetically changed glycine, serine, and betaine levels and insulin sensitivity levels as measured by fasting insulin or type 2 diabetes. However, for the clamp-based measures of insulin sensitivity we observed a suggestive association of a glycine-to-serine ratio–associated SNP, rs1107366 (near the ALDH1L1 gene). The allele that raises glycine-to-serine ratios increases clamp-based insulin sensitivity. The rs1107366–insulin sensitivity association needs further replication in a larger sample size, especially given that other glycine or serine-associated SNPs (e.g., rs715 and rs478093) are not associated with insulin sensitivity. The rs715 variant in CPS1, for example, explains a greater proportion of the variance in glycine levels (∼13% compared with ∼2% for rs1107366). It is also possible that rs1107366 could influence insulin sensitivity via non–glycine-mediated (pleiotropic) effects.

In an insulin-resistant state, the increase of hepatic gluconeogenesis would result in greater consumption of glycogenic amino acids, which may be accentuated in individuals with genetically influenced lower levels of these molecules and is consistent with the hypothesis of reverse causality (35). However, causal mechanisms in both directions remain plausible because gluconeogenesis is controlled in many different ways.

Our data highlight some candidate genes and protein products important in controlling circulating metabolite levels. At each locus, there is a clear candidate gene, although we cannot be certain which gene is affected by the associated SNP.

The SLC6A12 gene is a highly plausible candidate for influencing betaine levels. Betaine is an osmolyte used by cells for protection against hyperosmotic environments (36), and SLC6A12 encodes a highly conserved osmoregulator, which controls cellular volume by extrusion of betaine (37,38). Previous studies have shown that hyperosmolarity could induce insulin resistance by impairing insulin receptor substrate (IRS)-1 tyrosine phosphorylation and degradation of IRS1 and IRS2 in adipocytes (39,40). This connection may partly explain the association of betaine with insulin sensitivity.

The BHMT gene encodes a cytosolic enzyme that catalyses the conversion of betaine and homocysteine to dimethylglycine and methionine, respectively. The rs17823642 SNP is highly correlated with three SNPs associated with BHMT enzyme activity and protein level (rs41272270, r² = 1; rs16876512, r² = 0.925; and rs6875201, r² = 0.925) (41).

The previously described serine GWAS signal rs478093 is in PHGDH (also known as PGDH [see 3-PGDH in Fig. 2A]), the gene product of which catalyzes the first and rate-limiting step in the phosphorylation pathway of serine biosynthesis.

The rs715 SNP represents the same association signal as that identified between the SNP rs2216405 near CPS1 (r² = 0.47 with rs715) and glycine levels (31), but the variance in glycine levels explained by rs715 (12.87%) in our study compared with the variance explained by the rs2216405 SNP (8.64%) suggests that rs715 is a better marker for the causal variant. A recent study reported an association between rs715 and glycine levels specific to females, consistent with our results (32).

The enzyme encoded by CPS1 catalyzes synthesis of carbamoyl phosphate from ammonia and bicarbonate (Fig. 2B). Patients with defects in the function or expression of CPS1 suffer from life-threatening hyperammonemia (42). It is possible that variants at this locus perturb the conversion of ammonia and bicarbonate to carbamoyl phosphate. We hypothesize that CPS1 variants may cause excess ammonia, which may then lead to increased production of glycine and tetrahydrofolate in the glycine cleavage system (Figs. 2D and 3).

Glycine is a key component of the folate pathway (Fig. 3). The protein product of ALDH1L1 catalyzes the conversion of 10-formyltetrahydrofolate, NADP, and water to tetrahydrofolate, NADPH, and carbon dioxide (Fig. 2C). Our association between a SNP near the ALDH1L1 gene and glycine-to-serine ratio implicates this enzyme in glycine/serine conversion rate. This is in accordance with the knowledge that the glycine cleavage system, which accounts for ∼41% of whole-body glycine flux, is tightly linked with tetrahydrofolate in folate metabolism (43) (Fig. 2D and 3).

We observed a sex-specific association of CPS1 variants on glycine levels. This is consistent with the findings of Mittelstrass et al. (2011) (32). The variants at the CPS1 locus also have female-specific effects on homocysteine levels (33), and the glycine-raising allele is associated with raised homocysteine (44).

Glycine, serine, and betaine are linked to homocysteine and folate metabolism (44,45) (Fig. 3). The CPS1 variant rs715 is strongly correlated with rs7422339 (r² = 0.92), which was previously reported to be associated with homocysteine and folate levels (33). House, O’Connor, and Guenter (44) demonstrated that the plasma concentrations of homocysteine, glycine, and serine were all elevated in folate-deficient rats. From the link between betaine and homocysteine in the reaction catalyzed by BHMT, we hypothesized that if rs17823642 is associated with betaine level through reduced functioning of BHMT, then it would result in elevation of not only betaine but also homocysteine levels (46). We assessed the effect of rs17823642 on homocysteine levels in an independent European study (Invecchiare in Chianti, aging in the Chianti area [InCHIANTI]). We observed a nominal association in females (n = 575; β = 0.26 [95% CI 0.07–0.44]; P value 5.9 × 10⁻³) but not in males (n = 458; β = −0.02 [−0.25 to 0.21]; P value 0.87), where the betaine-raising allele also correlated with increased homocysteine levels. However, this association requires further confirmation with a larger sample size.

There are a number of limitations in our study. First, the triangulation approach for estimating expected effects does not take into account the complicated feedback mechanisms and interactions involved in controlling metabolite levels. The SNPs that we identified are associated with several metabolites (Table 4), which means that they do not provide specific instruments for one metabolite. However, the associations for any one SNP are all with metabolites closely connected in well-annotated pathways and therefore provide an instrument to test the relationship between alterations in those pathways and diabetes-related outcomes.

Second, our estimated effects are approximate, with metabolite levels only measured in RISC and observed estimates coming from separate studies. Nevertheless, the size of the MAGIC study provided very good power to see the expected very small effects on fasting insulin levels.

Finally, we have only been able to assess some of the many metabolites associated with insulin sensitivity. The metabolites selected in our study were those most strongly associated with clamp-measured insulin sensitivity in the RISC study (4), and we focused efforts on these metabolite traits. Some recent studies have reported other metabolites associated with dysglycemia (e.g., branched chain and aromatic amino acids) (3,5,47,48), but these metabolites have not been measured in RISC study.

In conclusion, our study provides novel insight into the genetic regulation of metabolite levels, particularly those involved in the glycine-related pathways closely correlated to insulin sensitivity. Genetic variants associated with metabolite levels provide an important approach to helping unravel the functional role of the metabolic pathways that influence diabetes-related traits.

This work was supported by grants from the Wellcome Trust (083270/Z/07/Z) and Medical Research Council (G0601261). The RISC study is partly supported by European Union Grant QLG1-CT-2001-01252. The EUGENE2 study is supported by a grant from the European Community’s FP6 EUGENE2 (LSHM-CT-2004-512013). J.W.K. is supported by American Heart Association National Fellowship to Faculty Transition Award 10FTF3360005. The RISC study is also supported by AstraZeneca.

W.E.G. is an employee of Metabolon, Inc. (Research Triangle Park, NC), as a Director in the Diagnostics division. Metabolon is a diagnostics and services company offering the industry’s leading biochemical profiling platform. No other potential conflicts of interest relevant to this article were reported.

W.X. designed the study, researched data, performed analyses, and cowrote the initial draft of the article. A.R.W. analyzed data. V.L. contributed data and reviewed and edited the manuscript. M.N.W. analyzed data. J.W.K. contributed data, contributed to discussion, and reviewed and edited the manuscript. S.A., T.L.A., T.Q., F.A., and J.P. contributed data and reviewed the manuscript. H.H. and T.H. provided samples and data from individual studies, contributed to discussion, and reviewed the manuscript. O.P. and U.S. provided samples and data from individual studies and reviewed the manuscript. M.L. contributed data and reviewed and edited the manuscript. J.M.D. provided samples and data from individual studies and reviewed the manuscript. J.J.N. performed phenotyping in individual studies and reviewed the manuscript. L.G. contributed to discussion and reviewed and edited the manuscript. E.F. and K.-P.A. provided samples and data from individual studies and reviewed the manuscript. W.E.G. performed phenotyping in individual studies, contributed to discussion, and reviewed the manuscript. T.M.F. designed the study, researched data, performed analyses, and cowrote the initial draft of the article. M.W. contributed data and reviewed and edited the manuscript. T.M.F. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

The authors thank Merck Research Laboratories for conducting DNA genotyping on the RISC samples and Stanford samples.